0 v0.28.4

• Retain all requested Ollama tool calls when they arrive in separate streaming chunks. Previously, later chunks overwrote earlier calls.
• Explain native tool-calling requirements when Ollama explicitly rejects the selected installed model/template, with /api/show guidance.

Validation: a two-chunk regression reproduces the original loss; all 19 Ollama runtime tests, full build and repository-wide TypeScript checks pass. No model inference or model downloads were needed for these protocol tests.

This fixes an independently reproduced protocol bug discovered while investigating a user report about Ollama + Qwen2.5 Coder. It does not establish that this was the cause of that user's failure, or that every installed model/template supports tools.